Variation and change in the use of hesitation markers in Germanic languages

Martijn Wieling1*, Jack Grieve2, Gosse Bouma1, Josef Fruehwald3, John Coleman4, and Mark

Liberman5

1University of Groningen, 2Aston University, 3University of Edinburgh, 4University of Oxford, 5University of Pennsylvania

In this study, we investigate cross-linguistic patterns in the alternation between UM, a

hesitation marker consisting of a neutral vowel followed by a final labial nasal, and UH, a

hesitation marker consisting of a neutral vowel in an open syllable. Based on a quantitative

analysis of a range of spoken and written corpora, we identify clear and consistent patterns of

change in the use of these forms in various Germanic languages (English, Dutch, German,

Norwegian, Danish, Faroese) and dialects (American English, British English), with the use of

UM increasing over time relative to the use of UH. We also find that this pattern of change is

generally led by women and more educated speakers. Finally, we propose a series of possible

explanations for this surprising change in hesitation marker usage that is currently taking

place across Germanic languages.

1. Introduction

Two basic hesitation markers (also referred to as fillers or filled pauses) are common in

modern Germanic languages: the UM form, which consists of a neutral vowel followed by a

final labial nasal, and the UH form, which consists of a neutral vowel in an open syllable. For

example, in the English language these forms are generally written as um and uh in American

English and as erm and er in British English. Similarly, in German a distinction is made

between in ähm or öhm and äh or öh, whereas in Dutch a distinction is made between ehm or

uhm and eh or uh. Similar forms appear to exist in all other Germanic languages. Hesitation markers, including UM and UH, have long been studied in linguistics,

primarily because their use has been seen as being directly related to the cognitive processes

responsible for the production of speech, specifically marking disfluencies (e.g., Maclay and

Osgood, 1959; Goldman-Eisler, 1968; Rochester, 1973; Crystal 1982; Levelt, 1983; Levelt &

Cutler, 1983; Schachter et al., 1991). For example, Schachter et al. (1991) found that lecturers

in the humanities used more hesitation markers than lecturers in the natural sciences when

teaching, but not when being interviewed. They argued that this difference is due to the larger

number of words from which a lecturer in the humanities must choose compared to a lecturer

in the natural sciences, where technical vocabulary is more strictly defined. Because

humanities lecturers have to make more decisions during speech production, they tend to use

more hesitation markers. In other words, hesitation markers are seen as marking disfluency

during language production. This general explanation for the use of hesitation markers has

been referred to as the symptom hypothesis (De Leeuw, 2007). Although disfluencies during language production would appear to explain many

occurrences of hesitation markers in spoken language, other explanations for the use of UM

and UH have been identified. For example, in a series of reaction time experiments, Brennan

and Schober (2001) found that hesitation markers were beneficial to comprehension, as

listeners were faster to select a target object after a filler was used in the stimulus sentence. Indeed, listeners appear to show a disfluency bias when encountering a hesitation marker. For

example, Arnold et al. (2013) found that when encountering disfluent speech, listeners were

more likely to expect a discourse-new referent. In line with this, Bosker et al. (2014) showed

that listeners were more likely to expect a low frequency as opposed to a high frequency word

after a disfluency marker, though listeners adapted this expectation on the basis of the speaker

(i.e. there was no higher expectation for a low-frequency word when listening to non-native

speakers). Similarly, Fox Tree (2001) showed that UH (but not UM) facilitated the speed with

which listeners were able to recognize upcoming words. Fraundorf and Watson (2011)

showed that hesitation markers improve recall whether or not they predict upcoming discourse

boundaries and that no such effect results from coughs of equal duration, ruling out a

processing time effect. In contrast to the symptom hypothesis, this type of explanation for the

usage of hesitation markers has been referred to as the signal hypothesis (De Leeuw, 2007).

Still other researchers have pointed out that UM and UH can be used to fulfill various

discursive functions (e.g. Swerts, 1998; Rendle-Short, 2004; Tottie, 2014). For example,

Swerts (1998) showed that hesitation markers can be used as markers of discourse structure,

with hesitation markers occurring more often with stronger discourse breaks than with weaker

discourse breaks. Similarly, Tottie (2014) argued that UM and UH can be used as discourse

markers, with a similar meaning as the discourse markers well and you know. Linguists have also directly compared the usage of UM and UH. For example, as noted

above, Fox Tree (2001) found that UH but not UM facilitated word recognition by listeners.

Alternatively, Shriberg (1994) reported that UM was more frequently found in sentence-initial

position than UH in American English, a result that Swerts (1998) replicated based on the

analysis of Dutch data. Similarly both Swerts (1998) and Clark and Fox Tree (2002) found

that UH tends to be used by speakers to mark minor delays, whereas UM tended to be used to

mark major delays. Such findings, however, have not been replicated by all researchers. For

example, O’Connell and Kowal (2005) argued that there are no functional differences in the

usage of UM and UH based on their analysis of six media interviews of Hillary Clinton.

Furthermore, based on a review of previous research, Corley and Stewart (2008) concluded

that there is no evidence that speakers have intentional control over the production of UM or

UH (see also Finlayson and Corley, 2012). Differences have also been found in the use of UM

and UH across Germanic languages. For example, De Leeuw (2007) reported that whereas

English and German speakers had a higher frequency of use of UM, Dutch speakers generally

had a higher frequency of use of UH.

The aforementioned studies have all focused on the different functions of UH and UM

from a structural perspective. However, researchers have also analyzed the effect of various

social factors on the choice between these two forms. For example, Rayson et al. (1997)

showed on the basis of a corpus analysis of the British National Corpus (BNC) that er (i.e.

UH) was the second-most characteristic word for male speech and the fourth-most

characteristic word for the speech of older (35+) speakers, whereas erm (i.e. UM) was the

ninth-most characteristic word for people from the upper social class, although they did not

directly contrast social patterns in the use of UM and UH. Liberman (2005), however, found

clear gender- and age-related patterns in the use of UH versus UM in corpora of transcribed

English-language telephone conversations (i.e. the Switchboard, Fisher Part 1 and Fisher Part

2 collections; Godfrey & Holliman, 1993; Cieri et al., 2004; Cieri et al., 2005). He observed

that the use of UH was higher for men than for women and for older speakers than for

younger speakers, whereas the use of UM was higher for women and younger speakers. In

other words, the frequency of UM relative to UH (i.e. the UM/UH ratio) was greater for

younger speakers and women. More recently, various other corpus-based studies have analyzed the use of hesitation

markers in English and have obtained similar results (see Tottie, 2011 for an overview). For

example, on the basis of two sub-corpora of the BNC (i.e. BNC-Demographic and BNC-

Context Governed), Tottie (2011) showed that women, younger people, and people from

higher socio-economic classes had a higher UM/UH ratio than men, older people and people

from lower socio-economic classes—a result that once again suggests that UM usage is rising

over time, led by women and speakers from higher classes. Similarly, Acton (2011) analyzed

the UM/UH ratio in American English based on the relatively recent Speed Dating Corpus

(SDC; Jurafsky et al., 2009) and the older Switchboard corpus (Godfrey and Holliman, 1993)

and obtained similar results, with women showing a greater UM/UH ratio than men in both

corpora. Based on the Switchboard corpus, Acton (2011) also showed that this pattern

persisted at the dialect-region level and when the gender of the hearer was taken into account

(i.e. same-gender dyads appeared to show a greater UM/UH ratio than different-gender

dyads). He also found that younger speakers had a greater UM/UH ratio than older speakers

and that the UM/UH ratio was greater for the more recent SDC than the Switchboard corpus

and therefore suggested that these results (together with the gender difference) might indicate

that a linguistic change is in progress. Similarly, Laserna et al. (2014) analyzed transcripts of

conversations collected by 263 American participants from five different studies (Mehl &

Pennebaker, 2003a, 2003b; Mehl, Gosling & Pennebaker, 2006; Fellows, 2009; Baddeley,

Pennebaker & Beevers, 2013), which were collected via electronically activated recorders

carried by the participants for two to three days, allowing for truly spontaneous conversations

to be obtained. Laserna et al. (2014) did not explicitly contrast the use of UM and UH in their

study, but they reported a significant correlation between gender (male: 1, female: 2) of r = -

.15 (p < .05) for UH, and r = -.09 (p > .05) for UM. Consequently, they concluded that

women showed a lower frequency of use for both UH and UM than men (since the correlation

coefficients are negative) (see also Bortfeld et al., 2001). However, as the reduction appears to

be greater for UH than UM, this result suggests that women in this study are characterized by

a greater UM/UH ratio than men. In addition, Laserna et al. (2014) reported a negative

correlation between age and UM use (r = -.21, p < .001), but not between age and UH use (r =

-.01, p > .05). As the use of UM (but not UH) decreases for older people, this implies that the

UM/UH ratio also decreases for older people, which once again implies that a change in

English hesitation marker usage is currently underway. Previous research on social variation in the use of UM and UH in British and American

English has thus repeatedly identified the same basic patterns: younger speakers and women

use relatively more UM than UH compared to older speakers and men (irrespective of the

potential categorical functional differences between the two alternatives). This type of pattern

is commonly identified in apparent-time sociolinguistic research and is seen as being

indicative of a linguistic change in progress (Labov, 1994) with the use of UM relative to UH

increasing over time. The apparent-time hypothesis assumes that most language is acquired

during childhood and remains relatively stable afterwards. Correspondingly, the speech of

older people is assumed to reflect the linguistic situation when these speakers were young.

Furthermore, variationist sociolinguistics studies have repeatedly found that language change

is led by women (e.g., see Labov, 2001). The first goal of this paper is therefore to assess

whether a change in hesitation markers usage is truly underway in the English language based

on detailed quantitative analyses of both longitudinal and apparent-time data. Furthermore,

because other Germanic languages have comparable hesitation markers, the second goal of

this paper is to investigate whether similar patterns of variation and change in the use of UM

and UH can be found in other Germanic languages, including Dutch, German, Norwegian,

Danish and Faroese.1

2. Data: Spoken language corpora

To compare patterns of linguistic variation and change in the use of the hesitation markers

UM and UH in Germanic languages, we analyzed a range of spoken language corpora

representing the English, Dutch, German, Norwegian, Danish and Faroese languages. For

1 While we focus on Germanic languages in this study, note that a similar gender-related pattern has been

recently observed in Mandarin speech (Yuan et al., submitted).

each of these corpora we generated a primary data set by extracting information about the

usage of UM and UH2 in the corpus as well as a range of social information about each

speaker.3 Most notably, we included gender and age. The age of the speakers may be used as a

way to assess linguistic change. This type of apparent time analysis is a common technique in

sociolinguistic research (see Labov, 1994) and is based on the assumption that if a change in

progress is taking place, younger speakers will tend to use the more modern form, whereas

older speakers tend to use the original form.

2.1. English For the English language, we analyzed five spoken language corpora, including three corpora

of American English, one corpus covering a wide range of British English dialects, and one

corpus of Scottish English. First, we analyzed the Switchboard Corpus of American English (Godfrey & Holliman,

1993), which contains data from approximately 2,400 two-sided telephone conversations

collected in 1990. We extracted all 91,001 tokens of UM (i.e. um) and UH (i.e. uh) from the

corpus, which were produced by a total of 520 different speakers. In addition, we recorded the

position (counted from the start of the utterance) and duration of the hesitation marker and the

duration of preceding and following pauses, as well as the age and gender of each speaker

(education level was not included), and the total number of words that they contributed to the

corpus. Second, we analyzed the Fisher Corpus of American English (Part 1 and Part 2) (Cieri

et al., 2004; Cieri et al., 2005), which contains transcripts of almost 12,000 telephone

conversations collected from 2002 to 2003. We extracted all 19,753 tokens of UM (i.e. um)

and UH (i.e. uh) from the corpus, which were produced by a total of 10,313 different

speakers. In addition, we obtained the age, gender and amount of education (in years) of each

speaker, and the total number of words that they contributed to the corpus.

Third, we analyzed the Philadelphia Neighborhood Corpus (PNC; Labov et al., 2013),

which contains transcripts of interviews with speakers from the Philadelphia area conducted

from 1973 to 2013. We extracted all 25,514 tokens of UM (i.e. um) and UH (i.e. uh) from the

corpus, which were produced by a total of 395 different speakers. In addition, we recorded the

duration of the hesitation marker, whether a pause occurred before or after the hesitation

marker, the year of recording, the age, gender and number of years of schooling of each

speaker, and the total number of words that they contributed to the corpus. Fourth, we analyzed the spoken component of the British National Corpus (BNC;

Coleman et al., 2012), which contains approximately seven million words recorded in 1993.

We extracted all 25,498 tokens of UM (i.e erm) and UH (i.e. er) from the corpus, which were

produced by a total of 960 different speakers. In addition, we recorded the duration of the

hesitation marker and the duration of the pause following the hesitation marker, as well as the

age and gender of each speaker, and the total number of words that they contributed to the

corpus. Fifth, we analyzed the HCRC Map Task Corpus of Scottish English (HCRC Map Task

Corpus, 1993), which contains transcribed speech collected from undergraduates at the

University of Glasgow in 1990, who were participating in a map task in which a guide had to

explain a route drawn on a paper map to a follower who only had a map without the route. We

2 Of course, transcribers may have made errors in assigning the label of the hesitation marker. However, it is

unlikely that these errors are specific to the gender and age of the speakers. 3 The data, methods and results associated with this analysis are available for download as supplementary

materials at the first author’s website (http://www.martijnwieling.nl) and at the Mind Research Repository

(http://openscience.uni-leipzig.de).

extracted all 1,987 tokens of UM (i.e. ehm, erm, mm4, um) and UH (i.e. eh, er, uh), which

were produced by a total of 64 different speakers (of which 61 subjects were Scottish). In

addition, we recorded the position of the hesitation marker in each utterance, as well as the

age, gender, and role (i.e. follower or guide) of each speaker, and the total number of words

that they contributed to the corpus.

2.2. Dutch For the Dutch language, we analyzed the Corpus Gesproken Nederlands (version 2.0) (CGN,

2006), which contains spoken transcribed speech (about 9 million words) from various

sources (e.g., spontaneous conversations, interviews, telephone dialogues) recorded from

1998 to 2004. We extracted all 228,619 tokens of UM (i.e. ehm, uhm) and UH (i.e. eh, uh)

from the corpus, which were produced by a total of 3,433 different speakers. In addition, we

recorded the position and duration of the hesitation marker, the duration of preceding and

following pauses, the preceding and following word, the part-of-speech tag of the preceding

and following word, as well as the age, gender, education level, nationality (Dutch, Belgian),

and level of preparedness (i.e. low for spontaneous speech, high for a televised speech) of

each speaker. Furthermore, we also extracted the total number of words that each speaker

contributed to the corpus.

2.3. German For the German language, we analyzed the Forschungs- und Lehrkorpus Gesprochenes

Deutsch (FLGD; Depperman, 2014), which contains about 100 hours of recorded speech

(about 1 million words) collected from 2005 to 2014. We extracted all 16,221 tokens of UM

(i.e. ähm, öhm) and UH (i.e. äh, öh), which were produced by a total of 238 different

speakers. In addition, we recorded the age and gender of each speaker.

2.4. Norwegian For the Norwegian language, we analyzed the Nordic Dialect Corpus and Syntax Database

(NDCSD; Johannessen et al., 2009), which contains approximately 2.8 million words from

conversations and interviews collected between 1951 and 2012. We extracted all 47,604

tokens of UM (i.e. em, EM, m, M, m-m, m_m) and UH (i.e. e, E, h-e) from the corpus that

were tagged as hesitation markers, which were produced by a total of 554 different speakers.

In addition, we recorded the year of recording, the age group (old: aged 50+, young: aged

between 18 and 30) and gender of each speaker, and the total number of words that they

contributed to the corpus.

2.5. Danish and Faroese Finally, for the Danish and Faroese languages, we analyzed the Faroese Danish Corpus

Hamburg (FADAC; Braunmüller, 2011), which contains 440,000 words collected on the

Faroe Islands from 2005 to 2009. We extracted all 4,504 tokens of UM (i.e. ehm, ehhm, eehm,

æhm, ææhm, øøhm, etc.) and UH (i.e. eh, ehh, eeh, æh, ææh, øøh, etc.) from the corpus,

which were produced by a total of 57 different speakers. In addition, we recorded the

language in which the interview was conducted (Danish, Faroese), the age and gender of each

speaker, and the total number of words that they contributed to the corpus.

4 We included mm, which accounts for 7.7% of all hesitation markers in this dataset, as it generally appears to be

used to mark hesitations, rather than for indicating assent (as opposed to mhm).

3. Data: Twitter corpora

In addition to analyzing various spoken language corpora, we also analyzed the use of UM

and UH in both American and Dutch Tweets, a written register that is especially informal and

shares several features with spontaneous speech. Notably, in instant message conversation, a

similar register of computer-mediated communication, Tagliamonte & Denis (2008) found

that the usage rates of discourse-pragmatic variables were broadly comparable to spoken

language corpora. Of course, the function of UH and UM will often be different in writing

than in speech, in large part because the use of UM or UH in written language is generally a

conscious process (i.e. it has to be typed) and Twitter is less interactional than spoken

language. This results in these forms being used primarily as discourse markers as opposed to

hesitation markers. For example, the following Twitter conversation shows that UM in

Twitter can be used to indicate irony, which appears to be far less common in spoken

language:

A: “Make fun of Jeb Bush's brother all you want, but he would've been dropping bombs

months ago.”

B: “um that's why everyone hates him”

Nevertheless, it is informative to test if patterns in the use of UM and UH can also be

observed in written language.

3.1. English For English Twitter, we analyzed a corpus of 6 billion words of American Tweets collected

by Diansheng Guo of the University of South Carolina in 2013, which only contains tweets

where the longitude and latitude of the user at the time of posting is known, as it was designed

for the analysis of geolinguistic variation. We extracted the 69,075 tokens of UM (i.e. um) and

UH (i.e. uh) from the corpus that were produced by the 25,852 users who contributed at least

1,000 total words to the corpus and whose username contained an unambiguous male or

female name (e.g. John2002 was designated as male, whereas Kate_1234 was designated as

female). Although this approach to identifying gender is not perfect, as some names will be

misclassified, we assume that the chances of misclassifications are relatively modest. In

addition, we recorded the gender of each user and the total number of words that they

contributed to the corpus. It would have also have been possible to use the username to

determine the year of birth, but very few users had a username containing a potential year of

birth (i.e. less than 1% of the 25,852 users).

3.2. Dutch For Dutch Twitter, we analyzed a corpus of 28.9 billion words of Dutch Tweets collected by

the Department of Information Science at the University of Groningen between 2011 and

2014. We extracted the 68,089 tokens of UM (i.e. uhm, um, euhm, ehm, etc.) and UH (i.e. uh,

uuh, eh, eeh, euh, etc.) from the corpus that were produced by the 38,651 users who

contributed at least 1,000 total words to the corpus and whose username contained an

unambiguous male or female name (as described above) and/or a four digit number ranging

between 1930 and 2009, which we used to estimate that user’s year of birth. In contrast to the

English dataset, the (much larger) Dutch Twitter dataset contained this four digit number

frequently in the usernames of Dutch Twitter users. This approach to identifying age also is

not perfect, as some names will be misclassified, but we assume that the chances of

misclassifications are relatively modest.

4. Analysis5

Because the dependent variable for each of the primary data sets is binary (i.e. the use of UM

versus UH or the number of tokens of UM versus the number of tokens of UH), we assessed

the effect of each of our predictor variables (e.g., age, gender, hesitation marker duration) on

the use of UM and UH using mixed-effects logistic regression (Agresti, 2007). By using

mixed-effects regression we are taking the structural variability associated with speakers into

account (see Baayen, 2008). This is important because some speakers may be more likely to

use UM (relative to UH) than others (i.e. modeled via a random intercept for speaker).

Similarly, the effect of each predictor may vary across speakers. For example, for some

speakers a longer duration of the pause following a hesitation marker may be more predictive

of the usage of UM than for other speakers. This would be modeled with a by-speaker random

slope for the duration of a following pause. Since we are using logistic regression, the

estimates need to be interpreted with respect to the logit scale (i.e. the logarithm of the odds of

observing UM rather than UH). Positive estimates indicate an increased probability of

observing UM together with increasing values of the predictor, whereas negative estimates

signal the opposite. An estimate of zero indicates that it has no effect on the probability of

observing UM.

For all of the primary data sets except one, we obtained the best-fitting model including

only significant predictors and supported random intercepts and random slopes. Predictors and

random intercepts and slopes were included if they reduced the Akaike Information Criterion

(AIC; Akaike, 1974) by at least 2, compared to the model without the random intercept or

slope (see also Wieling et al., 2014 for a similar approach). A reduced AIC indicates that the

additional complexity of the model is warranted given the increase in goodness of fit. Due to

the large number of predictors in the Dutch data set, however, we did not fit the best model

but rather fitted a random-intercepts-only model and assessed if the inclusion of individual

random slopes affected the significance of the predictors. We only included predictors that

remained significant in all cases in the final model. Given the large number of predictors in

this model, we also did not evaluate all possible interactions.

We assessed the goodness of fit of these models (including the random-effects structure)

by calculating the index of concordance C, which is known as the receiver operating

characteristic curve area ‘C’ (Harrell, 2001). Values of C greater than 0.8 indicate a successful

classifier, whereas a value of 0.5 indicates the classifier has no predictive power at all. All

models had C values close to or over 0.8 (see supplementary materials for exact values).

5. Results: Spoken language

Table 1 presents the effects (including associated estimations of effect size: the increase in

logits of the dependent variable for the categorical predictors, or per 1 standard deviation

increase of the numerical predictors) of the speaker-related predictors that were present in at

least two data sets (i.e. gender, age, education level, and year of recording) on the use of UM

over UH. Table 1 clearly shows that women are more likely than men to use UM as opposed

to UH across all data sets. Similarly, Table 1 shows that younger speakers are generally more

likely than older speakers to use UM as opposed to UH; only in the case of the relatively

small HCRC Corpus, does the effect of age not reach significance (p = 0.07). Table 1 also

5 Given that we analyzed nine independent data sets, we provide a simplified summary of the results for all

models together in this section, rather than reporting each individual model. The full details for each model can

be found in the supplementary materials (available at the Mind Research Repository: http://openscience.uni-

leipzig.de), which contains all data, all R commands used to generate the models, and all results for each

individual model, as well as detailed instructions on how to conduct the analysis.

shows that more or longer educated people are more likely to use UM as opposed to UH in the

Fisher corpus and the Dutch corpus, but that the effect of education in the PNC was non-

significant. In addition, the effect of education is much smaller than that of age. Finally, Table

1 shows that the use of UM over UH has increased over real-time in the PNC, the Norwegian

Corpus, and in the Dutch corpus. Figure 1 visualizes this result for the three data sets. For

each data set, the graph shows the proportion of UM over UH (i.e. UM/[UM+UH]) by year of

recording (divided into four groups containing roughly the same number of speakers) and

gender. The error bars indicate the 95% confidence interval (i.e. 1.96 standard errors below

and above the mean). It should be noted, however, that whereas the PNC (1973-2013) and the

Norwegian corpus (1951-2012) each span at least 40 years, the Dutch Corpus only spans 13

years and 90% of the data was recorded between 1999 and 2003. The effect of year of

recording is significant even while controlling for the age of the speaker (i.e. it is not an effect

of age grading; see Table 1). When year of recording is excluded from the analysis for the

PNC and instead only year of birth and age are taken into account, the most important

predictor clearly is year of birth; the effect of increasing age (i.e. older people are still more

likely to use UH) is only minimal (p = .04).

Significant interactions (e.g., between age and gender) were identified in some models;

however, because these interactions did not change the direction of the general effect (e.g., the

age effect was negative for both men and women, but less so for men than for women), we did

not explicitly include these interactions in Table 1 (see, however, supplemental materials for

the precise model specifications). Most important, these effects were found to be significant,

while controlling for the effect of other potential important predictors, such as the duration of

the pause before and after the hesitation marker (see Table 3, discussed below). Also note that

for the PNC (and for the Switchboard corpus, but not for the HCRC, nor the BNC), the

predictive value of the duration of the pause after the hesitation marker has diminished for

people born in more recent years (i.e. a longer pause is more likely to predict the occurrence

of UM over UH for older people than for younger people; see supplementary material). This

suggests, for these datasets, that younger people are using UM more across the board, and are

not simply more frequently signaling longer pauses.

Table 1. Effects of subject-related predictors on the choice of UM over UH for all data sets

Gender: Male

vs. Female Age:

Old vs. Young Education:

High/More vs.

Low/Less

Year of

Recording:

Increase vs.

Decrease

Switchboard F (1.03) Y (0.6z - 0.7z)

Fisher F (1.37) Y (0.39z) More (0.11z)

PNC F (1.31) Y (1.2z -1.7z) (More) (0.03z) Increase (0.54z)

BNC F (0.45) Y (0.45z)

HCRC F (2.30) (Y) (0.35z)

German F (0.43) Y (0.94z)

Norwegian F (0.23) Y (0.65) Increase (0.35z) Danish/Faroese F (0.59) Y (0.4z - 0.6z)

Dutch F (0.5 - 0.9) Y (0.3z - 0.6z) High (0.15z) Increase (0.09z) Significant (p < 0.05) and non-significant (category name put between parentheses) effects are listed; an empty cell indicates

the absence of that predictor in that data set. The values between parentheses indicate the effect size (in terms of logits: the

increase in probability of observing UM rather than UH) when the category changes to the one indicated or (when a

subscripted z is shown) when the value of the numerical predictor increases with 1 standard deviation. A range of values

indicates the predictor is involved in an interaction. In other words, the effect of age in the Switchboard corpus varies based

on the hesitation marker being phrase final (smaller effect) or not (larger effect), while the effect of gender and age varies per

country for the Dutch data set (larger for Belgium than for the Netherlands), and the effect of age varies per language in the

Danish/Faroese data set (larger for Faroese), and for gender in the PNC (larger for men).

Figure 2 presents four graphs for the American English Switchboard data set, which

visualize the relationship between age, gender and the use of UM and UH. The first graph

(top-left) plots the proportion of UM over UH (i.e. UM/[UM+UH]) by age (divided into four

age groups containing roughly the same number of speakers) and gender. Similarly as before,

the error bars indicate the 95% confidence interval (i.e. 1.96 standard errors below and above

the mean). This graph shows a clear increase in the proportion of UM over UH across age

groups for both men and women, with women consistently showing a higher rate of UM

usage than men. Note that all speakers in this corpus mostly use UH, with only women in the

two youngest age groups approaching 50% UM usage. The second graph (top-right) plots the

relative frequency of UM and UH taken together (i.e. total hesitation marker frequency

relative to all words in the corpus) by age and gender. This graph shows a clear decline in

hesitation marker usage across age groups for both men and women, with men consistently

using more hesitation markers than women (but note that this pattern is not observed in the

smaller HCRC and Danish/Faroese datasets, likely due to the large individual differences in

hesitation marker frequency; also Bell et al. (2000) found no gender differences in hesitation

marker usage for Swedish speakers). The third graph (bottom-left) charts the frequency of

UM relative to all words in the corpus by age and gender. This graph shows a clear increase in

UM use over age groups with women consistently using UM more frequently than men, even

though this gap appears to be closing in the youngest age group. Finally, the fourth graph

(bottom-right) plots the frequency of UH relative to all words in the corpus by age and

gender. This graph shows a clear decrease in UH usage across age groups with men

consistently using UH more frequently than women.

Figure 3 presents the same four graphs for the Dutch data set. Overall, the Dutch results

are similar to the American English results presented in Fig. 2. In particular, the first graph

(top-left) also shows a clear increase in the usage of UM over UH across age groups with

women showing a higher proportion of UM over UH than men, while the third graph (bottom-

left) shows a clear increase in the relative frequency of UM across age groups with women

using UM more often than men. Finally, the fourth graph (bottom-right) shows a decrease in

the relative frequency of UH across age groups, especially for women. Despite these

similarities, differences between the American English Switchboard data and the Dutch data

are apparent. Whereas hesitation markers in English have been showing a clear decrease in

frequency across age groups, the second graph (top-right) shows that there is no clear trend in

the overall usage of hesitation markers in Dutch (though the distinction between men and

women is similar). The visualizations for the other data sets, which can be found in the supplemental

material, all show relatively similar patterns. Most important, all data sets show an increase

across age groups in the use of UM over UH (with women having the highest proportion of

UM use) and an increase across age groups in the relative frequency of UM. In addition, most

data sets show a decrease across age groups in the use of UH. There are, however, differences

between the nine data sets. In particular, the relative frequency of hesitation markers across

age groups (i.e. the second graph in Figs 2 and 3) varies considerably across the nine data sets.

Despite generally following the same basic trends, there are also considerable

differences in the average overall proportions of UM over UH and the relative frequencies of

UM and UH across the nine data sets. These results are summarized in Table 2. For example,

the average proportion of UM over UH ranges from 27% to 64% for the five English corpora,

compared to 50% in the German corpus, 17% in the Danish corpus, 13% in the Norwegian

corpus, and 11% in the Dutch corpus.

Figure 1. Proportion of UM over UH for three data sets: PNC (top), Norwegian (middle) and

Dutch (bottom) by year of recording and gender.

Finally, Table 3 presents the effects (again including estimations of effect size) of the

hesitation marker-related predictors that were present in at least two data sets (i.e. the duration

of the hesitation marker, the duration or presence (for the PNC) of a pause before the

hesitation marker, the duration or presence (for the PNC) of a pause after the hesitation

marker, the presence of the hesitation marker at the start of the utterance, and the presence of

the hesitation marker at the end of the utterance) on the use of UM over UH. Table 3 only

presents results for the five data sets for which we were able to include information about the

duration and position of hesitation markers and pauses.

Figure 2. American English Switchboard data: proportion of UM over UH (top-left), relative

frequency of hesitation markers (top-right), relative frequency of UM (bottom-left), and

relative frequency of UH (bottom-right) by age and gender.

Table 2. Proportion of UM over UH and relative frequency of UM and UH for all data sets UM Proportion UM Relative

Frequency UH Relative

Frequency

Switchboard 0.2825 0.0075 0.0221

Fisher 0.6408 0.0099 0.0068 PNC 0.2765 0.0045 0.0132 BNC 0.4612 0.0043 0.0045

HCRC 0.5717 0.0081 0.0058 German 0.5017 (no word counts) (no word counts) Norwegian 0.1285 0.0026 0.0189 Danish/Faroese 0.1653 0.0020 0.0079 Dutch 0.1086 0.0037 0.0315

Figure 3. Dutch Spoken data: proportion of UM over UH (top-left), relative frequency of

hesitation (top-right), relative frequency of UM (bottom-left), and relative frequency of UH

(bottom-right) by age and gender.

Table 3. Effects of hesitation marker-related predictors on the choice of UM over UH

Duratio

n of

Marker

Duration/Presence

of pause before

Marker

Duration/Presence

of pause after

Marker

Initial Position

Final Position

Switchboard Longer

(0.87z) Longer (0.12z)

Longer (0.11z)

Initial

(0.67) Final

(1.06)

PNC Longer

(1.25z) (Absent) (-0.08)

Present / Longer (0.59) / (0.55z)

BNC Longer

(1.06z)

Longer (0.44z)

HCRC Initial

(0.83)

Final

(1.07)

Dutch Longer

(1.15z) Longer (0.17z)

Longer (0.47z)

Initial

(0.51) Final

(0.96) Significant (p < 0.05) and non-significant (category name put between parentheses) effects are listed; an empty cell indicates the absence of

that predictor in that data set. The values between parentheses indicate the effect size (in terms of logits: the increase in probability of observing UM rather than UH) when the category changes to the one indicated or (when a subscripted z is shown) when the value of the

numerical predictor increases with 1 standard deviation.

In general, all predictors showed positive estimates, indicating that higher values of the

predictors are associated with a greater likelihood of observing UM as opposed to UH.

Specifically, a longer duration (of the hesitation marker or the pause before or after the

hesitation marker) is associated with a greater likelihood of the hesitation marker being UM

rather than UH, while the occurrence of the hesitation marker in utterance-initial or utterance-

final position is also associated with a greater likelihood of the hesitation marker being UM

rather than UH. Note that in the case of the Philadelphia Neighborhood Corpus, the presence

of a pause before or after the hesitation marker is similar to the hesitation marker being

utterance initial or final as utterances were identified on the basis of the pauses (a pause of

200 ms. or more indicated the break between two utterances).

6. Results: Twitter

Figure 4 presents four graphs for the American English Twitter data set, which visualize the

relationship between gender and the use of UM and UH. The first graph (top-left) plots the

proportion of UM over UH and shows that women are more likely to use UM over UH than

men. The logistic mixed-effects regression model indicates this effect was significant (p <

.001). The second graph (top-right) plots the frequency of UM and UH taken together relative

to all words in the corpus and shows that women are more likely to use hesitation markers

than men. The third graph (bottom-left) plots the frequency of UM relative to all words in the

corpus and shows that women are more likely to use UM overall than men. The fourth graph

(bottom-right) plots the frequency of UH relative to all words in the corpus and shows that

women are more likely to use UH overall than men. These results for the proportion of UM

over UH and the relative frequency of UM agree with the results of the analysis of the

American English spoken language data sets (e.g., see Fig. 2); however, unlike the results of

the spoken analyses, women were found to have higher relative frequencies for UH and for

hesitation markers in general, likely reflecting functional differences in the use of UM and UH

in written language.

Figure 5 presents four graphs for the Dutch Twitter data set, which visualize the

relationship between age, gender and the use of UM and UH. The first graph plots the

proportion of UM over UH and shows that women and younger Twitter users are more likely

to use UM than men and older Twitter users, although in this case the youngest users were

found to reduce their use of UM compared to users from the second youngest group. The

logistic mixed-effects regression model indicates that the age effect was significant (p < .001)

but the gender effect was not (p = .13). However, note that a curvilinear pattern might be a

better fit to the data, with a decrease in proportion of UM over UH for the youngest users. The

second graph plots the frequency of UM and UH taken together relative to all words in the

corpus and shows that women and younger Twitter users are more likely to use hesitation

markers than men. Also in this case, the youngest users were found to reduce their use of

hesitation markers compared to users from the second youngest group. The third graph plots

the frequency of UM relative to all words in the corpus and shows that women and younger

Twitter users are more likely to use UM than men, although once again the youngest users

were found to reduce their use of UM compared to users from the second youngest group. The

fourth graph plots the frequency of UH relative to all words in the corpus and shows that

women and younger Twitter users are more likely to use UH than men, with a similar

deviating pattern for the youngest users. In terms of gender, these results (though not

significant) are in line with the results of the analysis of the American Twitter data. Although the results of the analysis of both the American and Dutch Twitter data

correspond reasonably well overall with the results of the analysis of the spoken language data

sets, the relative frequency of the hesitation markers in the Twitter data is an order of

magnitude lower than in the spoken language data, which likely reflects clear register

differences between speech and writing. Table 4 lists these values, for comparison with the

corresponding values for the spoken data sets presented in Table 2. Note that the proportion of

UM versus UH for the Dutch Twitter data is much larger than for the Dutch spoken data.

Again this is likely indicative of register differences between speech and writing.

Figure 4. American Twitter data: proportion of UM over UH (top-left), relative frequency of

UM and UH (top-right), relative frequency of UM (bottom-left), and relative frequency of UH

(bottom-right) by gender.

Table 4. Relative proportion of UM vs. UH and versus all words for the Twitter data sets

UM

Proportion UM Relative

Frequency UH Relative

Frequency

American English 0.5334 0.00025 0.00019 Dutch 0.6518 0.00011 0.00006

Figure 5. Dutch Twitter data: proportion of UM over UH (top-left), relative frequency of UM

and UH (top-right), relative frequency of UM (bottom-left), and relative frequency of UH

(bottom-right) by age and gender.

7. Discussion

The results of our analyses have shown that there is a consistent pattern of sociolinguistic

variation in the use of the hesitation markers UM and UH across many modern Germanic

languages. In English, Dutch, German, Norwegian, Danish and Faroese, UM is relatively

more common than UH in the language of women and younger speakers when compared to

the language of men and older speakers. Although gender and age patterns in the use of UM

and UH have been identified in previous research on the English language, this paper has

shown that this pattern holds across a wide variety of Germanic languages, as well as several

varieties of English (American, Scottish and other British dialects). Furthermore, because we

analyzed a wide variety of different corpora, this paper has also shown that this pattern is even

more pervasive, existing across a range of time periods and registers, including both speech

and writing. In addition to identifying a cross-linguistic pattern of language variation, the results of

our study strongly suggest that what has actually been identified is a cross-linguistic pattern of

language change. Because variation in the use of UM and UH shows a clear trend across age

groups, with younger speakers using UM rather than UH more often than older speakers, it

appears that there is a change in hesitation marker usage currently taking place across various

Germanic languages, with the use of UM rising over time. This type of apparent-time

evidence, which is common in sociolinguistic research (see Labov, 1994), is based on the

assumption that if a change is taking place, then younger speakers will generally be more

likely than older speakers to prefer the linguistic form that is on the rise. This interpretation of

our age-based results is strongly supported by our longitudinal analyses of the Philadelphian

English, the Norwegian, and (to a lesser extent) the Dutch corpora, which show that the use of

UM is rising in real time. Finally, our finding that women consistently use UM more often

than men is also consistent with this interpretation, as women have frequently been found to

lead linguistic change (see Labov, 1990). This study has therefore uncovered clear evidence

that a similar change is taking place in the use of the hesitation markers UM and UH

(irrespective of whether one accepts a categorical distinction between the two variants or not)

across a range of Germanic languages, with the use of UM as opposed to UH becoming more

frequent over time. This change in the use of hesitation markers is surprising because it is occurring

simultaneously across a relatively large and mostly mutually unintelligible set of Germanic

languages. Examples of cross-linguistic change are not well attested in the literature and it is

unclear how this type of change could have developed or could be maintained. Perhaps the

most basic question is whether this cross-linguistic change began in one language and then

spread to other languages, or whether it developed in all languages simultaneously. This is a

complex puzzle, one for which we cannot provide a definitive answer. In the remainder of this

paper, we therefore present a number of possible explanations for this cross-linguistic change,

discuss the strengths and weaknesses of each of these explanation, and consider how these

competing theories could be tested in future research.

One possible explanation for this cross-linguistic change is that there are independent

patterns of change in use of UM and UH occurring in all six of the Germanic languages for

various different reasons, which are coincidentally all moving in the same direction. Although

such an explanation is possible, completely independent changes progressing in unison across

six different languages is highly improbable. It is therefore necessary to consider other

hypotheses that directly explain why the same basic change is taking place across so many

Germanic languages. There would appear to be two general types of non-coincidental

explanations that could account for these results: the change may have spread through contact

from one of the languages to the others or a true parallel change may be taking place caused

by some factor that affects the use of these related hesitation markers across all the languages. Language contact is one possible explanation for this cross-linguistic change in

hesitation marker use. For example, lexical items in one language that refer to new concepts

are often borrowed into other languages that do not have words to refer to those concepts,

such as the English word ‘computer’. This word was borrowed into Dutch, German and

Danish, although not into Norwegian (datamaskin) or Faroese (telda). English forms, in

particular, would appear to be especially likely to spread through contact, because it is one of

the primary languages of mass media and the Internet, as well as being commonly used as a

second language by many speakers of other Germanic languages. Unfortunately, as all

datasets show an increase in UM use for younger speakers, and the age range covered in each

data set varies considerably, we are not able to identify when this potential development

would have started.

Even though it is well known that linguistic forms can spread through language contact,

which furthermore is often led by women (Van Ness, 1995), it is unclear if language contact

could explain the type of cross-linguistic change in hesitation marker usage identified in this

study. On the one hand, hesitation markers are relatively frequent in the English language,

ensuring that they would be present in the language to which non-native English speakers are

exposed. The proportion of UM (over UH) is also higher on average in the English language

corpora compared to the corpora for other Germanic languages, which is what we would

expect if the change originated in the English language. On the other hand, there is a

considerable range in the average usage of UM over UH in the English corpora (see Table 2),

which in some cases dips below the levels for German speakers in particular. The use of

hesitation markers would also generally appear to be a highly subconscious process and the

shift in usage of UM versus UH in the English language is a subtle change, only having been

identified here through a careful statistical analysis of large amounts of language data.

Furthermore, unlike the examples of language contact presented above, both forms involved

in this change already existed in all the Germanic languages under analysis, so that it is not

the specific form UM that would have spread but a pattern of change that affects a pre-

existing alternation. All of these factors presumably make it more difficult for variation in hesitation markers

to spread through contact than, for example, a new word that refers to a new concept.

However, perhaps that is what is happening here: UM might have taken on a new meaning or

function in English, and it is this meaning or function has spread through contact to other

Germanic languages, which already have a comparable form, To some extent we did control

for functional differences in the use of UM and UH by including various linguistic predictors

in our analyses. For example, UM tended to have a longer duration, was preceded and

followed by longer pauses, and was more frequently found at the beginning or end of an

utterance than UH. These results are in line with earlier studies (e.g., Clark and Fox Tree,

2002, Shriberg, 1994, Swerts, 1998), which found that UM is more likely to signal a major

delay (but see O’Connell and Kowal, 2005). Of course, a longer duration of UM is not

surprising, given that UM is essentially UH plus the labial nasal, but the gender and age-

related patterns still hold when these potential linguistic differences between the two

hesitation markers are controlled for. In addition, for most of the corpora analyzed here, the

overall relative frequency of UM and UH combined was found either to be decreasing or have

remained relatively steady over real or apparent time, which suggests that there has not been a

substantial increase in the use of UM or UH as discourse markers over this period of time. We

did not, however, analyze different linguistic functions of UM or UH. Most notably, as

discussed in the introduction, it is clear that hesitation markers can be used as discourse

markers, for instance to manage turn taking during a conversation, or to signal indecision,

disagreement, focus, or confusion. If UM, for example, is becoming more common as a

discourse marker over time compared to UH in English, then this change could explain the

rise of UM in English and could have been passed on to other Germanic languages through

contact. It should be noted, however, that relatively comparable age- and gender-related

patterns were found in the Twitter data, where UM and UH generally have different functions

than in spoken language. In addition to language contact, a cross-linguistic change could also be the result of

some linguistic or extra-linguistic process that causes each of the languages to change

independently but in parallel. For example, parallel changes can be a result of general

processes of sound change, such as elision, which involves the deletion of segments during

speech to facilitate articulation. There does not appear, however, to be any phonological

processes that would explain the rise in usage of UM compared to UH over time cross-

linguistically, such as a tendency for open syllables to close. In fact, the opposite is true: open

syllables are generally more common than closed syllables in languages of the world, and

furthermore syllables consisting solely of a vowel, such as UH, tend to develop onsets as

opposed to codas over time (Hyman, 2008). It also seems possible that UM could be reduced

to UH through elision in natural speech so as to accelerate language production. General

processes of phonological change therefore do not appear to explain the results of this study. Alternatively, a general extra-linguistic force could be responsible for a parallel change

in the usage of UM and UH across the six Germanic languages. For example, Biber et al.

(2010) found that noun phrase modification in English newspaper writing has become

syntactically more complex and compressed over time, and argue that this is due to the

increasing amount of information incorporated into newspapers in modern times and the

increasing use of word processing technology that has allowed reporters to devote more time

to carefully preparing and editing their texts. Similar societal changes could be affecting the

usage UM and UH cross-linguistically. For example, although there is general prescription

against using both forms in the English language (Erard, 2007), UM is arguably more polite

than UH (e.g., “polite yawning” is used to refer to yawning with the mouth closed; Hilgers et

al., 2000), given that UH leaves the mouth in an open position and that the UH sound is also

common reaction to physical pain, fatigue, sadness, and anger. Given the rise of living

standards, education level, mass media, and the service economy in the Western World over

the course of the 20th century, it is possible that people have become more self-conscious of

their language use, resulting in the rise of UM over UH across Germanic languages.

Unfortunately, the datasets analyzed in this study are not suitable for a more detailed

diachronic analysis of a potential shift towards more self-conscious language use.

Furthermore, to our knowledge there are no scientific studies describing a potential shift

towards (or away from) more self-conscious speech during the 20th century.

In conclusion, this study has shown that there is a clear change taking place across

modern Germanic languages, with UM rising in frequency relative to UH. Furthermore, we

have considered some possible explanations for this surprising cross-linguistic change, with

two hypotheses standing out as being most likely. The first explanation is that the change

originated in English and spread through contact with other Germanic languages, which have

similar forms, possibly reflecting semantic change in the use of UM, i.e. as a discourse

marker. The second explanation is that a parallel change is underway due to general societal

changes in communication in the Western World, for example with UM increasing in usage

because it is more self-conscious than UH. To assess these hypotheses both individually and

in conjunction, as well as potentially generating other explanations for the findings of this

study, it is necessary to conduct more detailed functional and social analyses of UM and UH

usage over time both within and across Germanic languages.

Acknowledgements

We thank Deborah Cameron, Michael Cysouw, Mark Dingemanse, Diansheng Guo, Daniel

Ezra Johnson, Alice Kasakoff, John Nerbonne, Andrea Nini, Emily Waibel, and three

anonymous reviewers whose comments have helped to improve this manuscript. Furthermore,

we thank Thomas Schmidt in helping us to extract the relevant information from the

Forschungs- und Lehrkorpus Gesprochenes Deutsch.

References

Acton, E. K. 2011. On gender differences in the distribution of um and uh. University of

Pennsylvania Working Papers in Linguistics 17(2), article 2.

Akaike, H. 1974. A new look at the statistical identification model. IEEE Transactions on

Automatic Control 19(6): 716-723.

Agresti, A. 2007. An Introduction to Categorical Data Analysis. John Wiley & Sons,

Hoboken, NJ, 2nd edition. Arnold, J. E., Fagnano, M., & Tanenhaus, M. K. 2003. Disfluencies signal theee, um, new

information. Journal of Psycholinguistic Research 32: 25-36. Baayen, R. H. 2008. Analyzing Linguistic Data: A Practical Introduction to Statistics using R.

Cambridge University Press.

Baddeley, J. L., Pennebaker, J. W., & Beevers, C. G. 2013. Everyday social behavior during a

major depressive episode. Social Psychological and Personality Science, 4: 445-452. Bamman, D., Eisenstein, J., & Schnoebelen, T. 2014. Gender identity and lexical variation in

social media. Journal of Sociolinguistics, 18(2): 135-160.

Biber D., Grieve J., & Iberri-Shea H. 2010. Noun phrase modification. In Rohdenburg G. &

Schlüter J. (eds.) One Language, Two Grammars? Differences between British and

American English, 182-193. Cambridge University Press.

Bell L., Eklund R., & Gustafson J. 2000. A Comparison of Disfluency Distribution in a

Unimodal and a Multimodal Human–Machine Interface. Proceedings of the International

Conference on Spoken Language Processing (ICSLP) 2000, 16–20 October 2000, Beijing,

China, vol. 3, 626-629.

Bortfeld, H., Leon, S.D., Bloom, J.E., Schober, M.F., & Brennan, S.E. 2001. Disfluency rates

in conversation: Effects of age, relationship, topic, role, and gender. Language and

Speech, 44: 123-147. Bosker, H. R., Quené, H., Sanders, T. J. M., & de Jong, N. H. 2014. Native 'um's elicit

prediction of low-frequency referents, but non-native 'um's do not. Journal of Memory

and Language, 75: 104-116.

Braunmüller, Kurt 2011. Faroese Danish Corpus Hamburg (FADAC Hamburg),

http://corpora.exmaralda.org/sfb_k8.html. Brennan, S. E., & Schober, M. F. 2001. How listeners compensate for disfluencies in

spontaneous speech. Journal of Memory and Language, 44(2): 274-296.

Cieri, C., Graff, D., Kimball, O., Miller, D., & Walker, K. 2004. Fisher English Training

Speech Part 1 Transcripts LDC2004T19. Web Download. Philadelphia: Linguistic Data

Consortium. Cieri, C., Graff, D., Kimball, O., Miller, D., & Walker, K. 2005. Fisher English Training Part

2, Transcripts LDC2005T19. Web Download. Philadelphia: Linguistic Data Consortium.

Clark, H. H., & Fox Tree, J.E. 2002. Using uh and um in spontaneous speaking. Cognition,

84: 73-111.

Coleman, J., Baghai-Ravary, L., Pybus, J., & Grau, S. 2012 Audio BNC: the audio edition of

the Spoken British National Corpus. Phonetics Laboratory, University of

Oxford, http://www.phon.ox.ac.uk/AudioBNC. Corley, M., & Stewart, O. W. 2008. Hesitation disfluencies in spontaneous speech: The

meaning of um. Language and Linguistics Compass, 2(4): 589-602. CGN. 2006. Corpus Gesproken Nederlands. Version 2.0.

http://tst-centrale.org/nl/producten/corpora/corpus-gesproken-nederlands/6-17 Crystal, D. 1982. Profiling linguistic disability. London: Arnold.

De Leeuw, E. 2007. Hesitation markers in English, German, and Dutch. Journal of Germanic

Linguistics 19(2): 85-114. Deppermann, A., & Schmidt, T. 2014. Gesprächsdatenbanken als methodisches Instrument

der Interaktionalen Linguistik - Eine exemplarische Untersuchung auf Basis des Korpus

FOLK in der Datenbank für Gesprochenes Deutsch (DGD2). In: Domke, Christine &

Gansel, Christa (eds.) Korpora in der Linguistik - Perspektiven und Positionen zu Daten

und Datenerhebung [=Mitteilungen des Deutschen Germanistenverbandes 1/2014], 4-17.

Erard, M. 2007. Um...Slips, Stumbles, and Verbal Blunders, and What They Mean. New York:

Pantheon Books.

Fellows, M. D. 2009. An exploration of emotion language use by preschool-aged children and

their parents: Naturalistic and lab settings. PhD thesis. The University of Texas at Austin Finlayson, I. R., & Corley, M. 2012. Disfluency in dialogue: an intentional signal from the

speaker? Psychonomic bulletin & review 19(5): 921-928.

Fox Tree, J. E. 2001. Listener’s uses of um and uh in speech comprehension. Memory and

Cognition 29: 320-326. Fraundorf, S.H., & Watson, D.G. 2011. The disfluent discourse: Effects of filled pauses on

recall. Journal of Memory and Language 65(2): 161-175. Godfrey, J., & Holliman, E. 1993. Switchboard-1 Release 2 LDC97S62. DVD. Philadelphia:

Linguistic Data Consortium. Goldman-Eisler, F. 1968. Psycholinguistics: Experiments in spontaneous speech. London:

Academic Press. HCRC Map Task Corpus 1993. LDC93S12. Web Download. Philadelphia: Linguistic Data

Consortium.

Hilgers F. J., van Dam F. S., Keyzers S., Koster M. N., van As C. J., & Muller M. J. 2000.

Rehabilitation of olfaction after laryngectomy by means of a nasal airflow-inducing

maneuver: the polite yawning technique. Archives of Otolaryngology–Head & Neck

Surgery 126(6): 726-732.

Hyman, L. M. 2008. Universals in phonology. The linguistic review 25(1-2): 83-137.

Johannessen, J.B., Priestley, J., Hagen, K., Åfarli, T. A., & Vangsnes, Ø. A. 2009. The Nordic

Dialect Corpus - an advanced research tool. In: Jokinen, Kristiina and Eckhard Bick

(eds.) Proceedings of the 17th Nordic Conference of Computational Linguistics

NODALIDA. NEALT Proceedings Series Volume 4, 73-80. Jurafsky, D., Ranganath, R., & McFarland, D. 2009. Extracting social meaning: Identifying

interactional style in spoken conversation. In: Proceedings of Human Language

Technologies: The 2009 Annual Conference of the North American Chapter of the

Association for Computational Linguistics. ACL, 638–646.

Labov, W. 1990. The intersection of sex and social class in the course of linguistic change.

Language Variation and Change 2(2): 205-254. Labov, W. 1994. Principles of Language Change. Volume 1: Internal Factors. Oxford:

Wiley-Blackwell. Labov, W. 2001. Principles of Language Change. Volume 2: Social Factors. Oxford: Wiley-

Blackwell. Labov, W., Rosenfelder, I., & Fruehwald, J. 2013. One hundred years of sound change in

Philadelphia: Linear incrementation, reversal, and reanalysis. Language 89(1): 30-65. Laserna, C. M., Seih, Y. T., & Pennebaker, J. W. 2014. Um... who like says you know. Filler

word use as a function of age, gender, and personality. Journal of Language and Social

Psychology 33(3): 328-338.

Levelt, W. J. M. 1983. Monitoring and self-repair in speech. Cognition 14: 41-104. Levelt, W. J. M., & Cutler, A. 1983. Prosodic marking in speech repair. Journal of Semantics

2: 205-217. Liberman, M.(2005. Young men talk like old women.

http://itre.cis.upenn.edu/~myl/languagelog/archives/002629.html (first accessed 6

November 2005). Maclay, H., & Osgood, C. E. 1959. Hesitation phenomena in spontaneous English speech.

Word 15: 19-44.

Mehl, M. R., Gosling, S. D., & Pennebaker, J. W. 2006. Personality in its natural habitat:

Manifestations and implicit folk theories of personality in daily life. Journal of

Personality and Social Psychology 90: 862-877. Mehl, M. R., & Pennebaker, J. W. 2003a. The social dynamics of a cultural upheaval: Social

interactions surrounding September 11, 2001. Psychological Science 14: 579-585. Mehl, M. R., & Pennebaker, J. W. 2003b. The sounds of social life: A psychometric analysis

of students’ daily social environments and natural conversations. Journal of Personality

and Social Psychology 84: 857-870.

O’Connell, D. C., & Kowal, S. 2005. Uh and um revisited: Are they interjections for signaling

delay? Journal of Psycholinguistic Research 34: 555-576. Rayson, P., Leech, G., & Hodges, M. 1997. Social differentiation in the use of English

vocabulary: Some analyses of the conversational component of the British National

Corpus. International Journal of Corpus Linguistics 2: 133-152.

Rendle-Short, J. 2004. Showing structure: Using um in the academic seminar.

Pragmatics, 14(4), 479-498. Rochester, S. R. 1973. The significance of pauses in spontaneous speech. Journal of

Psycholinguistic Research 2(1): 51-81. Schachter, S., Christenfeld, N., Ravina, B., & Bilous, F. 1991. Speech disfluency and the

structure of knowledge. Journal of Personality and Social Psychology 60(3): 362-367. Shriberg, E. E. (1994). Preliminaries to a theory of speech disfluencies. PhD dissertation,

University of California, Berkeley. Swerts, M. 1998. Filled pauses as markers of discourse structure. Journal of pragmatics

30(4): 485-496.

Tagliamonte, S. A., & Denis, D. 2008. Linguistic Ruin? Lol! Instant Messaging and Teen

Language. American Speech 83(1): 3-34. doi:10.1215/00031283-2008-001

Tottie, G. 2011. Uh and um as sociolinguistic markers in British English. International

Journal of Corpus Linguistics 16(2): 173-197. Tottie, G. 2014. On the use of uh and um in American English. Functions of Language 21(1):

6-29.

Van Ness, S. 1995. Ohio Amish women in the vanguard of a language change: Pennsylvania

German in Ohio. American speech 70(1): 69-80.

Yuan J., Xiaoying X., Lai W., & Liberman M. Submitted. Pauses and pause fillers in

Mandarin monologue speech: The effects of sex and proficiency.

Wieling, M., Montemagni, S., Nerbonne, J., & Baayen, R. H. 2014. Lexical differences

between Tuscan dialects and standard Italian: Accounting for geographic and socio-

demographic variation using generalized additive mixed modeling. Language 90(3): 669-

692.